The performance of deep learning models suffers the domain shift between the training dataset and test dataset frequently. Domain adaptation is a popular machine learning technique to tackle it. Generally, existing domain adaptation methods learn domain-invariant features and seldom consider class-level matching. To address it, we propose a Cooperative Attention Generative Adversarial Network (CAGAN) by generating verisimilar target samples with given class labels and implementing class-level transfer. Specifically, we integrate Coupled Generative Adversarial Networks (CoGAN) into a classification network. The shared generator fed with class semantic codes steers downstream generators to produce source and target samples with the same high-level semantics. However, the single weight-sharing mechanism cannot guarantee the semantic consistency of generated sample pairs in an enormous domain gap, so we propose a semantic-consistent loss to reduce the domain shift in the shared generative space. In addition, attention layers with adaptive factors are proposed to embed into the shared generator, contributing to capturing more suitable representations of both domains. Extensive experiments demonstrate that our proposed model can achieve the best or comparable results on several standard domain adaptation benchmarks.